Multi-agent cooperation for particle accelerator control

摘要

Quite exhaustive theoretical studies exist in DAI, but it seems that there is not enough feedback from practice. In this paper we present practical investigations for applying and justifying the theoretical DAI results in a real industrial controls environment, and, conversely, we discuss the theoretical aspects of practical findings in these applied investigations made for accelerator control and operation. The results presented here are partly based on the research carried out at CERN during the ESPRIT-II Project ARCHON™. A generalized hypothesis is introduced, based on a unified view of control, monitoring, diagnosis, maintenance and repair tasks leading to a general method of cooperation for expert systems by exchanging hypotheses and leads to a mapping of different tasks in accelerator control to cooperating agents. This has been tested for task and result-sharing cooperation scenarios and we also report on studies carried out in relation to accelerator timings and their diagnosis as well as the transformation of these systems into a multi-agent community. Generalized hypotheses also allow us to treat the repetitive diagnosis-recovery cycle as a task-sharing cooperation. Problems with such a loop or even recursive calls between the different agents are discussed.